1 / 22

Transaction Management: Concurrency Control

Transaction Management: Concurrency Control. CS634 Class 19, Apr 14, 2014. Slides based on “Database Management Systems” 3 rd ed , Ramakrishnan and Gehrke. Database. Files. Pages. Tuples. Multiple-Granularity Locks. Hard to decide what granularity to lock tuples vs. pages vs. files

otis
Download Presentation

Transaction Management: Concurrency Control

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Transaction Management: Concurrency Control CS634Class 19, Apr 14, 2014 • Slides based on “Database Management Systems” 3rded, Ramakrishnan and Gehrke

  2. Database Files Pages Tuples Multiple-Granularity Locks • Hard to decide what granularity to lock • tuples vs. pages vs. files • Shouldn’t have to decide! • Data containers are nested: contains

  3. IS IX S X -- Ö Ö Ö Ö Ö -- IS Ö Ö Ö Ö IX Ö Ö Ö S Ö Ö Ö Ö X New Lock Modes, Protocol • Allow transactions to lock at each level, but with a special protocol using new intention locks • Before locking an item, must set intention locks on ancestors • For unlock, go from specific to general (i.e., bottom-up). • SIX mode: Like S & IX at the same time.

  4. Multiple Granularity Lock Protocol • Each transaction starts from the root of the hierarchy • To get S or IS lock on a node, must hold IS or IX on parent node • To get X or IX or SIX on a node, must hold IX or SIX on parent node. • Must release locks in bottom-up order

  5. Isolation Levels in Practice • Databases default to RC, read-committed, so many apps run that way, can have their read data changed, and phantoms • Web apps (JEE, anyway) have a hard time overriding RC, so most are running at RC • The 2PL locking scheme we studied was for RR, repeatable read: transaction takes long term read and write locks • Long term = until commit of that transaction

  6. Read Committed (RC) Isolation • 2PL can be modified for RC: take long-term write locks but not long term read locks • Reads are atomic as operations, but that’s it • Lost updates can happen in RC: system takes 2PC locks only for the write operations: R1(A)R2(A)W2(B)C2W1(B)C1 R1(A)R2(A)X2(B)W2(B)C2X1(B)W1(B)C1 (RC isolation) • Update statements are atomic, so that case of read-then-write is safe even at RC • Update T set A = A + 100 (safe at RC isolation) • Remember to use update when possible!

  7. Snapshot Isolation (SI) • Multiversion Concurrency Control Mechanism (MVCC) • This means the database holds more than one value for a data item at the same time • Used in PostgreSQL (open source), Oracle, others • Readers never conflict with writers unlike traditional DBMS (e.g., IBM DB2)! Read-only transactions run fast. • Does not guarantee “real” serializability • But: ANSI “serializability” fulfilled, i.e., avoids anomalies in the ANSI table • Found in use at Microsoft in 1993, published as example of MVCC

  8. Snapshot Isolation - Basic Idea: • Every transaction reads from its own snapshot (copy) of the database (will be created when the transaction starts, or reconstructed from the undo log). • Writes are collected into a writeset (WS), not visible to concurrent transactions. • Two transactions are considered to be concurrent if one starts (takes a snapshot) while the other is in progress.

  9. First Committer Wins Rule of SI • At the commit time of a transaction its WS is compared to those of concurrent committed transactions. • If there is no conflict (overlapping), then the WS can be applied to stable storage and is visible to transactions that begin afterwards. • However, if there is a conflict with the WS of a concurrent, already committed transaction, then the transaction must be aborted. • That’s the “First Committer Wins Rule“ • Actually Oracle uses first updater wins, basically same idea, but doesn’t require separate WS

  10. Write Skew Anomaly of SI • In MVCC, data items need subscripts to say which version is being considered • Zero version: original database value • T1 writes new value of X, X1 • T2 writes new value of Y, Y2 • Write skew anomaly schedule: R1(X0) R2(X0) R1(Y0) R2(Y0,) W1(X1) C1 W2(Y2) C2 • Writesets WS(T1) = {X}, WS(T2) = {Y}, do not overlap, so both commit. • So what’s wrong—where’s the anomaly?

  11. Write Skew Anomaly of SI R1(X0) R2(X0) R1(Y0) R2(Y0,) W1(X1) C1 W2(Y2) C2 • Scenario: • X = husband’s balance, orig 100, • Y = wife’s balance, orig 100. • Bank allows withdrawals up to combined balance • Rule: X + Y >= 0 • Both withdraw 150, thinking OK, end up with -50 and -50. • Easy to make this happen in Oracle at “Serializable” isolation. • See conflicts, cycle in PG, can’t happen with full 2PL • Can happen with RC/locking

  12. How can an Oracle app handle this? • If X+Y >= 0 is needed as a constraint, it can be “materialized” as sum in another column value. • Old program: R(X)R(X-spouse)W(X)C • New program: R(X)R(X-spouse) W(sum) W(X)C • So schedule will have W(sum) in both transactions, and sum will be in both Writesets, so second committer aborts. • Or, after the W(X), run a query for the sum and abort if it’s negative.

  13. Oracle, Postgres: new failure to handle • Recall deadlock-abort handling: retry the aborted transaction • With SI, get "can't serialize access“ • ORA-08177: can't serialize access for this transaction • Means another transaction won for a contended write • App handles this error like deadlock-abort: just retry transaction, up to a few times • This only happens when you set serializable isolation level

  14. Other anomalies under SI • Oldest sailors example • Both concurrent transactions see original sailor data in snapshots, plus own updates • Updates are on different rows, so both commit • Neither sees the other’s update • So not serializable: one should see one update, other should see two updates. • Task Registry example: • Both concurrent transactions see original state with 6 hours available for Joe • Both insert new task for Joe • Inserts involve different rows, so both commit

  15. Fixing the task registry • Following the idea of the simple write skew, we can materialize the constraint “workhours <= 8” • Add a workhours column to worker table • Old program: • if sum(hours-for-x)+newhours<=8 • insert new task • New program: • if workhours-for-x + newhours <=8 • { update worker set workhours = workhours + newhours… • insert new task • }

  16. Fixing the Oldest sailor example • If the oldest sailor is important to the app, materialize it! Create table oldestsailor (rating int primary key, sidint)

  17. Oracle Read Committed Isolation • READ COMMITTED is the default isolation level for both Oracle and PostgreSQL. • A new snapshot is taken for every issued SQL statement (every statement sees the latest committed values). • If a transaction T2 running in READ COMMITTED mode tries to update a row which was already updated by a concurrent transaction T1, then T2 gets blocked until T1 has either committed or aborted • Nearly same as 2PL/RC, though all reads occur effectively at the same time for the statement.

  18. Transaction Management: Crash Recovery CS634 • Slides based on “Database Management Systems” 3rded, Ramakrishnan and Gehrke

  19. ACID Properties Transaction Management must fulfill four requirements: • Atomicity: either all actions within a transaction are carried out, or none is • Only actions of committed transactions must be visible • Consistency: concurrent execution must leave DBMS in consistent state • Isolation: each transaction is protected from effects of other concurrent transactions • Net effect is that of some sequential execution • Durability: once a transaction commits, DBMS changes will persist • Conversely, if a transaction aborts/is aborted, there are no effects

  20. Recovery Manager • Crash recovery • Ensure that atomicity is preserved if the system crashes while one or more transactions are still incomplete • Main idea is to keep a log of operations; every action is logged before its page updates reach disk (Write-Ahead Log or WAL) • The Recovery Managerguarantees Atomicity & Durability

  21. Motivation • Atomicity: • Transactions may abort – must rollback their actions • Durability: • What if DBMS stops running – e.g., power failure? • Desired Behavior after system restarts: • T1, T2 & T3 should be durable • T4 & T5should be aborted (effects not seen) crash! T1 T2 T3 T4 T5

  22. Assumptions • Concurrency control is in effect • Strict 2PL • Updates are happening “in place” • Data overwritten on (deleted from) the disk • A simple scheme is needed • A protocol that is too complex is difficult to implement • Performance is also an important issue

More Related